I want to talk more about Large Language Models (LLMs) and ChatGPT, as it’s all anyone asks me about when I give talks, either in Europe or here in the States. It doesn’t matter where. It’s always ChatGPT. Not self-driving cars. Not robotics. It’s the tech that Sam Altman dissed as “cloning human speech” that has apparently captured everyone’s attention. If I don’t talk about it, I’m not talking about AI. Got it! So I’ll talk about it.

Garden Pathing AI

Not to go all Altman on everyone, but I think LLMs are nothing but a “garden path” technology. Let me explain. In linguistics, a garden path sentence is one that starts out grammatically, but leads the reader to a dead-end. The meaning is either radically different from the original idea of where the sentence is going, or it’s simply nonsensical. Example: “The complex houses married and single soldiers and their families.” 

Cool. What? What happened? This: we want to interpret “complex houses” as a noun phrase (like “the blue butterfly,” where “blue” is an adjective describing the noun “butterfly”) but then the rest of the sentence is nonsensical. Taken as a noun phrase, “complex houses” can’t “marry.” There’s a subject-verb disagreement that makes it nonsensical (most subject-verb disagreements confuse tense, but we can make sense of the sentence anyway. Not here.) What the sentence really means is this: the “complex” is the noun and “houses” is to be treated like a verb. Ahhh. Now it makes sense.  The complex houses three hundred soldiers and a terrified physicist. Easier.

Garden path sentences are fun, because among other delights, they force your brain to actually PARSE and UNDERSTAND what’s being said. It was Grice, I think, who articulated the common sense axioms of communication, and one of them was (I’m paraphrasing) that we apply a principle of charity in communication, and assume that someone has a point. Garden path sentences force us to make the “axiom” explicit, because we keep trying to make sense of a garden path sentence until we get the point (some of them actually are nonsensical, which makes it more confounding and difficult still).

Large Language Models are a kind of garden path in AI. What I mean is, we get this great simulated understanding, but to do it we had to spend billions of dollars, hijack 25,000 plus computers (with GPUs), and pretend like the input isn’t simply human writing on the web. But this is just the preamble. What makes LLMs truly a garden path for AI is simply that, without an actual understanding of what’s being said, it’s a flat out dead-end for making legitimate progress in the field. It teaches us nothing, basically. We can’t get systems like ChatGPT to stop confabulating or hallucinating, because we can’t understand their vast numerical (not linguistic) complexity in the first place. Imagine a pocket calculator (as in, the olden timey days) that performed arithmetic perfectly, except that every one thousand or two thousand calculations, it got the answer wrong. How do you commercialize that? Hope it doesn’t screw anything up (it’s mostly accurate)?

Whenever I make this point some true believer invariably complains that humans screw stuff up too. Yes, we do. But the cases are not analogous. In our own case, we can trace back the reasoning steps and at least in principle clear things up. “Oh, you didn’t hear about X yet. Well, it affects your results here.” And so on. We simply can’t do this post-mortem on LLMs or systems like ChatGPT. We get what they give us, and if we don’t have preexisting knowledge, we’re down the rabbit hole of untruth. The cases are simply not the same, and the difference is one that matters. A lot. LLMs and ChatGPT will happily dig in their heels and defend nonsense. Yes, we know people like this, but defending a contradiction or just talking b.s. would take an obstinate and stupid human being indeed.

Yeah, we did it again. We wandered down a beautiful garden path made possible by deep neural networks and later the 2017 innovation, the “attention” mechanism, and we ended up realizing we’ve dead-ended AI for a generation or more. Since no one is, apparently, working on anything else these days, we can be guaranteed that the moment of achieving AGI, let alone superintelligence, is further away than ever. Maybe that’s why they initially fired Altman. He said as much. 

And so it is with AI research and development since its inception. It’s a history of picking problems already computational in nature, or problems that admit of a computational treatment. Sometimes funders go rogue, and ask for non-computational results: I once sat on a Zoom call with DARPA officials (who shall remain nameless) sincerely asking how we can include “consciousness” in our systems. I’ll get right on that. But mostly, we solve problems with AI systems by making the world look more like something a computer can solve. Or stealing text made by humans on the web to pass it off as ChatGPT smarts. Bye, broader visions. Bye, humanism.

How Does That Make You feel?

Natural language interpretation and generation was always a big dream, and from early on. Many AI enthusiasts (or sceptics) remember Joseph Wizenbaum’s ELIZA chatbot. It answered questions with questions, like a Rogerian psychotherapist:

Woman: Men are all alike.

ELIZA: In what way?
Woman: They’re always bugging us about something or other.
ELIZA: Can you think of a specific example?
Woman: Well, my boyfriend made me come here.
ELIZA: Your boyfriend made you come here?
Woman: He says, I’m depressed much of the time.
ELIZA: I’m sorry to hear that you’re depressed.
Woman: It’s true, I am unhappy.
ELIZA: Do you think coming here will help you not to be unhappy?

This goes on and on. Weizenbaum noticed that human participants would start emotionally connecting, even bonding with the chatbot. They would insist it was helping them (maybe it was). But his point — and later writings suggest he developed ELIZA tongue in cheek, as a kind of mild rebuttal to AI overreach — was simply that the system has no clue of the emotions or feelings or intellect of its human interlocutors. It’s just inserting your comment into another comment, checking some grammar, and appending a “?” at the end of its response. That’s it. The whole thing was a sham. Or, a garden pathing of AI, into yet another beguiling and exciting dead-end.

AGI is Nowhere in Sight

Final thought. I have friends and colleagues who’ve gone almost religious about LLMs and conversational systems like ChatGPT. They tell me I’m a hopeless sceptic. They tell me it’s true AGI. I disagree. I think Microsoft, who has invested thirteen billion and counting into developing the technology, and with its new “Copilot,” intends to use the tech to vastly improve office productivity — think the old Microsoft Word paperclip helper, but with better answers — will not recoup its cash investment. I think it’s an impressive but fundamentally flawed approach. I think it’s a garden pathing of AI. And how long will it take before we start doing real research?

Time will tell. I think it won’t take that much longer to see what’s really going on.

Originally published at Colligo.